A novel systolic array structure for DCT

This paper presents a new algorithm for the implementation of discrete cosine transform (DCT), based on the idea of reformulating prime N-length DCT into two cyclic convolutions with exactly the same structure, which are implemented with a proposed fast cyclic convolution-based systolic array structure. The proposed algorithm can save (N-1)/2 multiplications and 2Nregisters, at the cost of only (N-1)/2 additions of those used in previous designs. The I/O is kept low because of the simple control complexity of the algorithm. Furthermore, this new algorithm preserves all the other benefits of very large-scale integration algorithms based on circular correlation or cyclic convolution, such as regular and simple structure.